A Survey of Weak Instruments

and Weak Identification in Generalized

Method of Moments
James H. Stock
Harvard University and the National Bureau of Economic Research, Cambridge, MA 02138

Jonathan H. Wright
Federal Reserve Board, Washington, DC 20551

Motohiro Yogo
Harvard University, Cambridge, MA 02138

Weak instruments arise when the instruments in linear instrumental variables (IV) regression are weakly
correlated with the included endogenous variables. In generalized method of moments (GMM), more
generally, weak instruments correspond to weak identification of some or all of the unknown parameters.
Weak identification leads to GMM statistics with nonnormal distributions, even in large samples, so that
conventional IV or GMM inferences are misleading. Fortunately, various procedures are now available
for detecting and handling weak instruments in the linear IV model and, to a lesser degree, in nonlinear
GMM.

KEY WORDS: Instrument relevance; Instrumental variables; Similar tests.

1. INTRODUCTION 2. Empirical researchers often confront weak instruments.

A subtle but important contribution of Hansen and Single-
ton’s (1982) and Hansen’s (1982) original work on general-
ized method of moments (GMM) estimation was to recast the
requirements for instrument exogeneity. In the linear simul-
taneous equations framework then prevalent, instruments are
exogenous if they are excluded from the equation of inter-
est; in GMM, instruments are exogenous if they satisfy a
conditional mean restriction that, in Hansen and Singleton’s
(1982) application, was implied directly by a tightly speci-
fied economic model. Although these two requirements are the
same mathematically, they have conceptually different start-
ing points. The shift from debatable [“incredible,” according
to Sims (1980)] exclusion restrictions to first-order conditions
derived from economic theory has been productive, and care-
ful consideration of instrument exogeneity is now a standard
part of solid empirical analysis using GMM.
But instrument exogeneity is only one of the two criteria
necessary for an instrument to be valid. Recently, the other
criterion—instrument relevance—has received increased atten-
tion by theoretical and applied researchers. It now appears
that some, perhaps many, applications of GMM and instru-
mental variables (IV) regression have what is known as “weak
instruments” or “weak identification,” that is, instruments that
are only weakly correlated with the included endogenous
variables. Unfortunately, weak instruments pose considerable
challenges to inference using GMM and IV methods.
This survey of weak instruments and weak identification
has five themes:
Finding exogenous instruments is hard work, and the features
that make an instrument plausibly exogenous, such as occur-
ring sufficiently far in the past to satisfy a first-order condi-
tion or the as-if random coincidence that lies behind a quasi-
experiment, can also work to make the instrument weak.
3. It is not useful to think of weak instruments as a “small-
sample” problem: Bound, Jaeger, and Baker (1995) provided
an empirical example of weak instruments despite having
329,000 observations.
4. There are methods more robust to weak instruments than
conventional GMM.
5. What to do about weak identification is a more diffi-
cult issue in nonlinear GMM than in linear IV regression, and
much theoretical work remains.
This survey emphasizes the linear IV regression model with
homoscedastic, serially uncorrelated errors, mainly because
much more is known about weak instruments in this case.
Section 2 provides a primer on weak instruments in linear
IV regression, and Section 3 discusses some empirical appli-
cations that confront weak instruments. Sections 4–6 discuss
recent econometric methods for handling weak instruments in
the linear model with homoscedastic errors: detection of weak
instruments (Sec. 4); methods that are fully robust to weak
instruments, at least in large samples (Sec. 5); and partially
robust methods that are somewhat simpler to use (Sec. 6).
Section 7 turns to weak identification in GMM for nonlinear
models and/or heteroscedastic or serially correlated errors.
Section 8 concludes.

1. If instruments are weak, then the sampling distributions

© 2002 American Statistical Association
of GMM and IV statistics are in general nonnormal, and stan- Journal of Business & Economic Statistics
dard GMM and IV point estimates, hypothesis tests, and con- October 2002, Vol. 20, No. 4
fidence intervals are unreliable. DOI 10.1198/073500102288618658

518
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM
Although many of the key ideas of weak instruments have
been understood for decades, most of the literature on solu-
tions to the problem of weak instruments is quite recent, and
this literature is expanding rapidly. We both fear and hope
that much of the practical advice in this survey will soon be
outdated.
2. A PRIMER ON WEAK INSTRUMENTS
IN THE LINEAR REGRESSION MODEL
Many of the problems posed by weak instruments in the
linear IV regression model are best explained in the context
of the classical version of that model with fixed exogenous
variables and iid normal errors. We therefore begin by using
this model to show how weak instruments lead to the two-
stage least squares (TSLS) estimator with a nonnormal sam-
pling distribution, regardless of sample size. In general, exact
distributions of IV statistics are not a practical basis for infer-
ence, and the section concludes with a synopsis of asymptotic
methods designed to retain the insights gained from the finite-
sample distribution theory.
2.1 The Linear Gaussian Instrumental Variables
Regression Model With a Single Regressor
The linear IV regression model with a single endogenous
regressor and no included exogenous variables is
y = Y + u (1)
and
Y = Z + v (2)
where y and Y are T × 1 vectors of observations on endoge-
nous variables, Z is a T × K matrix of instruments, and u
and v are T × 1 vectors of disturbance terms. The instru-
ments are assumed to be nonrandom (fixed). The errors ut vt
 Because u = Eu
v +  = uv /v2 v +  with  and v inde-
t = 1     T  are assumed to be iid N 0 , where the ele-
v = 0 and the result follows.
ments of  are u2 , uv , and v2 , and let  = uv /u v . Equa-
tion (1) is the structural equation, and  is the scalar parameter
of interest. The reduced-form equation (2) relates the endoge-
nous regressor to the instruments.
2.1.3. The Concentration Parameter. The concentration
parameter, 2 , is a unitless measure of the strength of the
instruments and is defined as
2 =  Z  Z/v2  (3)
A useful interpretation of 2 is in terms of F , the F statistic
for testing the hypothesis  = 0 in (2) (i.e., the “first-stage
F statistic”). Let F
be the infeasible counterpart of F , com-
puted using the true value of v2 . Then K F
is distributed as
a noncentral chi-squared distribution with degrees of freedom
K and noncentrality parameter 2 , and EF
 = 2 /K + 1.
If the sample size is large, then F and F
are close, so
EF   2 /K + 1. Thus larger values of 2 /K shift out the
distribution of the first-stage F statistic, and F − 1 can be
thought of as an estimator of 2 /K.
519
2.1.4. An Expression for the Two-Stage Least Squares
Estimator. The TSLS estimator is ˆ TSLS = Y  PZ y/
Y  PZ Y , where PZ = ZZ  Z−1 Z  . Rothenberg (1984) pre-
sented a useful expression for the TSLS estimator that
obtains by substituting Y  PZ u =  Z  u + v PZ u and Y  PZ Y =
 Z  Z + 2 Z  v + v PZ v into the expression for ˆ TSLS − 
and collecting terms,
zu + Suv /
ˆ TSLS −  = u /v   (4)
1 + 2zv / + Svv /2
√
where zu =  Z  u/u  Z  Z, zv =  Z  v/v ×
√
 Z  Z, Suv = v PZ u/v u , and Svv = v PZ v/v2 .
Under the assumptions of fixed instruments and normal errors,
zu and zv are standard normal random variables with correla-
tion , and Suv and Svv are quadratic forms of normal random
variables with respect to the idempotent matrix PZ .
Because the distributions of zu , zv , Suv , and Svv do not
depend on the sample size T , the sample size enters the dis-
tribution of ˆ TSLS only through the concentration parameter.
If 2 is small, then the terms zv , Suv , and Svv in (4) lead to a
nonnormal distribution. In contrast, the leading term zu dom-
inates if 2 is large, yielding the usual normal approximation
to the distribution of ˆ TSLS . Formally, 2 plays the role in (4)
usually played by the number of observations: As 2 becomes
large, the distribution of ˆ TSLS −  is increasingly well
approximated by the N 0 u2 /v2  distribution. For the normal
approximation to the distribution of the TSLS estimator to be
accurate, the concentration parameter must be large.
2.1.5. Bias of the Two-Stage Least Squares Estimator in
the Unidentified Case. When 2 = 0 (equivalently, when
 = 0), the instruments are not just weak, but irrelevant. In this
case, the mean of the TSLS estimator is the probability limit of
the ordinary least squares (OLS) estimator, plim(ˆ OLS ). Specif-
ically, when K ≥ 3 so that its mean exists, Eˆ TSLS −  =
plimˆ OLS − = uY /Y2 . To derive this result, note that when
 = 0, ˆ TSLS −  = v PZ u/v PZ v uv = uY , and v2 = Y2 .
pendent, Ev PZ 
When the instruments are relevant but weak, the TSLS esti-
mator is biased toward plim(ˆ OLS ). Specifically, define the
“relative bias” of TSLS to be the bias of TSLS relative to the
inconsistency of OLS, that is, Eˆ TSLS − /plimˆ OLS − .
When 2 is moderately large, the TSLS relative bias is approx-
imately inversely proportional to 2 /K − 2, a result that
holds even if the errors are not normally distributed (Buse
1992).
2.1.6. Numerical Examples. Figures 1(a) and 1(b) show
the pdf’s of the TSLS estimator and its t statistic for var-
ious values of the concentration parameter when the true
value of  is 0. The other parameter values mirror those
of Nelson and Startz (1990a, b): K = 1, u = v = 1, and
 = 99, so plimˆ OLS  = 99. For small values of 2 /K, such
as Nelson and Startz’s value of .25, the distributions are strik-
ingly nonnormal, even bimodal. As 2 /K increases, the dis-
tributions approach the usual normal limit.
The dramatic Nelson–Startz results drew econometricians’
attention to the problem of weak instruments. Their results
build on a large literature on the exact distribution of IV
520
Figure 1. pdf of TSLS Estimator (a) and t Statistic (b) for 2 /K =
021252102100; One Instrument (K = 1); and  = 199, Computed by Monte
Carlo Simulation.
estimators under the assumptions of fixed instruments and
iid normal errors (e.g., Sawa 1969; Richardson 1968). How-
ever, the results in this literature, comprehensively reviewed
by Phillips (1984), are offputting and pose substantial compu-
Journal of Business & Economic Statistics, October 2002
quality of the approximations when the sample is large but
2 /K is not.
2.2.1. Edgeworth Expansions. An Edgeworth expansion
is a representation of√the distribution of the statistic of inter-
est in powers of 1/ T . As Rothenberg (1984) pointed out
in the fixed-instrument,
√ normal-error model, an Edgeworth
expansion in 1/ T with a fixed model is formally equivalent
to an Edgeworth expansion in 1/. In this sense, Edgeworth
expansions improve on the conventional normal approximation
when  is small enough for the term in 1/2 to matter, but not
so small that the terms in 1/3 and higher matter. Rothenberg
(1984) suggested that the Edgeworth approximation is “excel-
lent” for 2 > 50 and “adequate” for 2 as small as 10, as
long as the number of instruments is small (less than ).
2.2.2. Many-Instrument Asymptotics. Although the prob-
lems of many instruments and weak instruments might at first
seem different, they are in fact related. With many strong
instruments, the adjusted R2 of the first-stage regression would
be nearly 1, so a small first-stage adjusted R2 indicates that
the instruments, taken as a set, are weak. Bekker (1994) for-
malized this notion by developing asymptotic approximations
for a sequence of models with fixed instruments and normal
errors, in which the number of instruments, K, is proportional
to the sample size and 2 /K converges to a constant, finite
limit; similar approaches were taken by Anderson (1976),
Kunitomo (1980), and Morimune (1983). Many-instrument
asymptotic distributions are generally normal, and simulation
evidence suggests that these approximations are good for both
moderate and large values of K, although they cannot cap-
ture the nonnormality evident in the Nelson–Startz example of
Figure 1. Distributions derived using this approach generally
depend on the distribution of the errors (see Bekker and van
der Ploeg 1999), so some procedures that are justified using
many-instrument asymptotics require adjustments for nonnor-
mal errors. However, rate and consistency results are more
robust to nonnormality (see Chao and Swanson 2002).
2.2.3. Weak-Instrument Asymptotics. Like many-instru-
ment asymptotics, weak-instrument asymptotics (Staiger and
Stock 1997) involves a sequence of models chosen to keep
tational challenges. Moreover, the assumptions of fixed instru-
ments and normal errors are generally too restrictive to be
appropriate in empirical application. To overcome these lim-
itations, researchers have used asymptotic approximations, to
which we now turn.
2.2 Asymptotic Approximations
Conventional asymptotic approximations to finite-sample
distributions are calculated for a fixed model in the limit that
2 /K constant as T → . However, unlike many-instrument
asymptotics, K is held fixed. Technically, the sequence of
models considered is the same as used to derive the local
asymptotic power of the first-stage F test √ (a “Pitman drift”
parameterization in which  is in a 1/ T neighborhood of
0). Staiger and Stock (1997) showed that under general condi-
tions on the errors and with random instruments, many results
that hold exactly in the fixed-instrument, normal-error model
can be reinterpreted as holding asymptotically, with simplifi-
cations arising from the consistency of Z  Z/T and of the esti-
T → , but sometimes this approach does not provide the mator for v2 .
most useful approximating distribution. This is the case for
the weak instruments problem; as is evident in Figure 1, the
usual fixed-model asymptotic normal approximations can be 3. EMPIRICAL EXAMPLES
quite poor when the concentration parameter is small, even if
3.1 Estimating the Returns to Education
the number of observations is large. For this reason, alterna-
tive asymptotic methods are used to analyze IV statistics in the In an influential article, Angrist and Krueger (1991) pro-
presence of weak instruments. Three such methods are Edge- posed using the quarter of birth as an instrument to circumvent
worth expansions, many-instrument asymptotics, and weak- ability bias in estimating the returns to education. The date of
instrument asymptotics. These methods aim to improve the birth, they argued, should be uncorrelated with ability, so that
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM
quarter of birth is exogenous; because of mandatory school-
ing laws, quarter of birth should also be relevant. With large
samples from the U.S. census, they estimated the returns to
education by TSLS, using as instruments quarter of birth and
its interactions with state and year of birth binary variables,
for as many as 178 instruments.
Surprisingly, despite the large number of observations
(329,000 or more), the instruments, taken together, are weak
in some of the Angrist–Krueger regressions. This point was
first made by Bound et al. (1995), who provided Monte Carlo
results showing that in some specifications, similar point esti-
mates and standard errors obtain if each individual’s true quar-
ter of birth is replaced by a randomly generated quarter of
birth. Because the results with the randomly generated quarter
of birth must be spurious, this suggests that the results with
the true quarter of birth are misleading. The source of these
misleading inferences is weak instruments; in some specifica-
tions, the first-stage F statistic is less than 2, suggesting that
2 /K might be 1 or less (recall that EF  − 1  2 /K. In
these specifications, there are a few strong instruments (the
quarter of birth binary variables) and many weak ones (their
interactions with state and year), resulting in a combined set
of instruments that is weak. An important conclusion is that it
is not helpful to think of weak instruments as a “finite-sample”
problem that can be ignored if one has many observations.
3.2 The Log-Linearized Euler Equation in the
Consumption-Based Capital Asset-Pricing Model
The first empirical application of GMM was Hansen and
Singleton’s (1982) investigation of the consumption-based
capital asset pricing model (CCAPM). In its log-linearized
form, the first-order condition of the CCAPM with constant
relative risk aversion can be written as
Ert+1 + ! − "#ct+1 
Zt
= 0 (5) to the case of included exogenous regressors by replacing y, Y ,
where " is the coefficient of relative risk aversion (here also
the inverse of the elasticity of intertemporal substitution),
#ct+1 is the growth rate of consumption, rt+1 is the log gross
return on some asset, ! is a constant, and Zt is a vector of
variables in the information set at time t (Hansen and Single-
ton 1983; Campbell 2001 for a survey).
The coefficients of (5) can be estimated by GMM using
Zt as an instrument. One way to proceed is to use TSLS
with rt+1 as the dependent variable; another is to apply TSLS
with #ct+1 as the dependent variable; and a third is to use
a method, such as limited-information maximum likelihood
(LIML), that is invariant to the normalization. Under stan-
dard fixed-model asymptotics, these estimators are asymptoti-
cally equivalent, so it should not matter which method is used.
However, as discussed in detail by Neely, Roy, and Whiteman
(2001) and Yogo (2002), this does matter greatly in practice,
with point estimates of " ranging from small (Hansen and
Singleton 1982, 1983) to very large (Hall 1988; Campbell and
Mankiw 1989).
The first-stage F statistics in these regressions are fre-
quently less than 5 (Yogo 2002), and it appears that weak
instruments can explain many of these seemingly contradic-
tory results (Stock and Wright 2000; Neely et al. 2001). For
521
an instrument to be strong, it must be a good predictor of
either consumption growth or an asset return, depending on
the normalization, but both are notoriously difficult to predict.
So finding weak instruments in this application should not be
a surprise.
3.3 Macroeconometric Examples
Weak identification can also be a concern in GMM esti-
mation of macroeconomic equations with expectational terms.
For example, Ma (2002) and Mavroeidis (2001) suggested
that weak instruments can be an issue in GMM estimation of
the hybrid New Keynesian Phillips curve (Fuhrer and Moore
1995; Gali and Gertler 1999). Other macroeconomic appli-
cations that confront weak identification include estimates
of New Keynesian output equations (Fuhrer and Rudebusch
2002) and some structural vector autoregressions (Pagan and
Robertson 1998).
4. DETECTION OF WEAK INSTRUMENTS
This section discusses methods for detecting weak instru-
ments. In general, the linear IV regression model has n
endogenous regressors, so that Y and v in (2) are T × n. The
methods for detecting weak instruments (and the definition of
the concentration parameter) depend on n. We first discuss
inference based on the first-stage F statistic when there is a
single endogenous regressor, then turn to the case of n > 1.
The section concludes with an alternative approach to infer-
ence about weak instruments proposed by Hahn and Hausman
(2002).
To keep things simple, the formulas in Sections 4–6 apply
to the case in which there are no included exogenous regres-
sors. These formulas and methods, however, generally extend
and Z by the residuals from their projection onto the included
exogenous regressors and by modifying the degrees of free-
dom as needed. Unless noted otherwise, the methods discussed
in Sections 4–6 do not require fixed instruments and normally
distributed errors for their asymptotic justification.
4.1 The First-Stage F Statistic
Before discussing how to use the first-stage F statistic to
detect weak instruments, we need to provide a precise defini-
tion of weak instruments.
4.1.1. A Definition of Weak Instruments. A practical
approach is to define a set of instruments to be weak if 2 /K
is small enough that inferences based on conventional normal
approximating distributions are misleading. In this approach,
the definition of weak instruments depends on the purpose to
which the instruments are put, combined with the researcher’s
tolerance for departures from the usual standards of inference
(i.e., bias, size of tests). For example, suppose that one is using
TSLS and want its bias to be small. Accordingly, one measure
of whether a set of instruments is strong is whether 2 /K is
sufficiently large so that the TSLS relative bias (as defined in
Sec. 2) is at most (say) 10%; if not, then the instruments are
522 Journal of Business & Economic Statistics, October 2002

deemed weak. Alternatively, if interested in hypothesis test- 4.2 Extension of the First-Stage F Statistic to n > 1
ing, one could define instruments to be strong if 2 /K is large
When there are multiple endogenous regressors, the concen-
enough that a 5% hypothesis test rejects no more than (say) 1/2 1/2
tration parameter is a K × K matrix, V V  Z ZV V , where
15% of the time; otherwise, the instruments are weak. These V V is the covariance matrix of the vector of errors vt . To
two definitions—one based on relative bias and the other based avoid introducing new notation, we refer to the concentra-
on size—in general yield different threshold values of 2 /K; tion parameter as 2 in both the scalar and matrix cases. The
thus instruments might be weak if used for one application, quality of the usual normal approximation is governed by the
but not if used for another. matrix 2 /K. Because the predicted values of Y from the first-
Here we consider the two definitions of weak instruments in stage regression can be highly correlated, for the usual normal
the previous paragraph: The TSLS relative bias could exceed approximations to be good, it is not sufficient that some ele-
10%, or the actual size of the nominal 5% TSLS t test could ments of 2 /K are large. Rather, the matrix 2 /K must be
exceed 15%. As shown by Stock and Yogo (2001), under large in the sense that its smallest eigenvalue is large.
weak-instrument asymptotics, each of these definitions implies From a statistical perspective, when n > 1, the n first-stage
a threshold value of 2 /K. If the actual value of 2 /K exceeds F statistics are not sufficient for the concentration matrix even
this threshold, then the instruments are strong (e.g., TSLS rel- with fixed regressors and normal errors (see Shea 1997 for a
ative bias is <10%). Otherwise, the instruments are weak. discussion). Instead, inference about 2 can be based on the
4.1.2. Ascertaining Whether Instruments Are Weak Using n × n matrix analog of the first-stage F statistic,
the First-Stage F Statistic. In the fixed-instrument, normal-
GT = 
 V V Y  PZ Y 
−1/2 −1/2
V V /K (6)
error model, or, alternatively, under weak-instrument asymp-
totics, the distribution of the first-stage F statistic depends
where V V = Y  MZ Y /T − K, MZ = I − PZ , and I is a con-
only on 2 /K and K. Hence the F statistic is useful for mak- formable identity matrix. Under weak-instrument asymptotics,
ing inference about 2 /K. As Hall, Rudebusch, and Wilcox EGT  → 2 /K +I. Cragg and Donald (1993) proposed using
(1996) showed in Monte Carlo simulations, simply using F GT to test for partial identification (cf. Choi and Phillips
to test the hypothesis of nonidentification  = 0 is an 1992)—specifically, testing the hypothesis that the matrix 
inadequate screen for problems caused by weak instruments. has rank L against the alternative that it has rank greater than
Instead, we follow Stock and Yogo (2001) and use F to test L, where L < n. From the perspective of IV inference, mere
the null hypothesis that 2 /K is less than or equal to the weak- instrument relevance is insufficient; instead, the instruments
instrument threshold against the alternative that it exceeds the must be strong in the sense that 2 /K is large. Accordingly,
threshold. Stock and Yogo (2001) considered the problem of testing the
For selected values of K, Table 1 reports weak-instrument null hypothesis that a set of instruments is weak against the
threshold values of 2 /K and critical values of F for testing alternative that they are strong, where instruments are defined
the null hypothesis that instruments are weak. For example, to be strong if conventional TSLS inference is reliable for
under the TSLS relative bias definition of weak instruments, any linear combination of the coefficients. By focusing on the
if K = 5, then the threshold value of 2 /K is 5.82, and the worst-behaved linear combination, this approach is conserva-
test that 2 /K ≤ 582 rejects in favor of the alternative that tive but tractable, and Stock and Yogo provided tables of crit-
2 /K > 582 if F ≥ 1083. Evidently the first-stage F statistic ical values, analogous to those in Table 1, based on the mini-
mum eigenvalue of GT .
must be large, typically exceeding 10, for TSLS inference to
be reliable.
4.3 A Test of the Null of Strong Instruments
The methods discussed so far have been tests of the null of
Table 1. Selected Critical Values for Weak Instrument Tests for TSLS
weak instruments. Hahn and Hausman (2002) reversed the null
Based on the First-stage F statistic and alternative and proposed a test of the null that the instru-
ments are strong against the alternative that they are weak.
Relative bias > 10% Actual size of 5% test > 15% They noted that when there is a single endogenous regressor
Number of
instruments Threshold F statistic 5% Threshold F statistic 5% n = 1 and the instruments are strong, normalization of the
K 2 /K critical value 2 /K critical value regression (the choice of dependent variable) should not mat-
ter. Thus the TSLS estimator in the forward regression of y
1 1
82 8
96
2 4
62 11
59 on Y and the inverse of the TSLS estimator in the reverse
3 3
71 9
08 6
36 12
83 regression of Y on y are asymptotically equivalent [to order
5 5
82 10
83 9
20 15
09 op T −1/2 ] with strong instruments, but this is not the case if
10 7
41 11
49 15
55 20
88
15 7
94 11
51 21
69 26
80 the instruments are weak. Accordingly, Hahn and Hausman
(2002) developed a statistic comparing the forward and reverse
NOTE: The second column contains the smallest values of 2 /K that ensure that the bias regression estimators (and their extensions when n = 2). They
of TSLS is no more than 10% of the inconsistency of OLS. The third column contains the 5%
critical values applicable when the first-stage F statistic is used to test the null that 2 /K is less suggested that if this statistic rejects the null hypothesis, then
than or equal to the value in the second column against the alternative that 2 /K exceeds that a researcher should conclude that his or her instruments are
value. The final two columns present the analogous weak-instrument thresholds and critical
values when weak instruments are defined so that the usual nominal 5% TSLS t test of the weak. Otherwise, the researcher can treat the instruments as
hypothesis  = 0 has size potentially exceeding 15%. (Source: Stock and Yogo 2001.) strong.
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM
5. FULLY ROBUST INFERENCE
WITH WEAK INSTRUMENTS
This section discusses hypothesis tests and confidence sets
for  that are fully robust to weak instruments, in the sense
that these procedures have the correct size or coverage rates
regardless of the value of 2 (including 2 = 0) when the sam-
ple size is large (specifically, under weak-instrument asymp-
totics). We focus on the case where n = 1, but these methods
generalize to joint inference about  when n > 1.
Several fully robust tests have been proposed; consistent
with earlier Monte Carlo studies, the results here suggest that
none appears to dominate the others. Moreira (2001) provided
a theoretical explanation of this in the context of the fixed-
instrument, normal-error model. In that model, there is no uni-
formly most powerful test of the hypothesis  = 0 , a result
that also holds more generally under weak-instrument asymp-
totics. In this light, the various fully robust procedures repre-
sent trade-offs, with some working better than others, depend-
ing on the true parameter values.
5.1 A Family of Fully Robust Gaussian Tests
Moreira (2001) considered the system (1) and (2) with fixed
instruments and normally distributed errors. Suppose that the
reduced-form equation for y is y = Z + w. Let / denote
the covariance matrix of the reduced-form errors, wt vt
 , and
for now suppose that / is known. We are interested in testing
the hypothesis  = 0 .
Moreira (2001) showed that under these assumptions, the
statistics
Z  Z−1/2 Z  Y b0 Z  Z−1/2 Z  Y /−1 a0

=   and  =   (7)
b0 /b0 a0 /−1 a0
are sufficient for  and , where Y = y Y
 b0 = 1 − 0
 , K0  =  (9)
and a0 = 0 1
 . Thus for the purpose of testing  = 0 , it
suffices to consider test statistics that are functions of only

and  , say g
  . Moreover, under the null hypothesis  =
0 , the distribution of  depends on , but the distribution
of
does not; thus, under the null hypothesis,  is sufficient
for . It follows that a test of  = 0 based on g
   is
similar if its critical value is computed from the conditional
distribution of g
   given  . Moreira (2001) also derived
an infeasible power envelope for similar tests under the further
assumption that  is known. In practice,  is not known;
when K > 1, feasible tests cannot achieve the power envelope,
and there is no uniformly most powerful test of  = 0 .
In practice, / is unknown, so the statistics in (7) cannot
be computed. However, under weak-instrument asymptotics,
/ can be estimated consistently under the null and, moreover,
the results in the preceding paragraph generalize to stochastic
instruments and nonnormal errors. Accordingly, let
 and 
denote
and  evaluated with /  = Y  MZ Y /T −K replacing
/, where MZ = I − PZ . We refer to Moreira’s (2001) family
of tests, based on statistics of the form g
  , as Gaussian
similar tests.
5.2 Three Gaussian Similar Tests
We now turn to three Gaussian similar tests: the Anderson–
Rubin (AR) statistic, Kleibergen statistic, and Moreira statistic.
523
5.2.1. The Anderson–Rubin Statistic. More than 50 years
ago, Anderson and Rubin (1949) proposed testing the null
hypothesis  = 0 using the statistic
y − Y0  PZ y − Y0 /K


AR0  = =  (8)
y − Y0  MZ y − Y0 /T − K K
One definition of the LIML estimator is that it minimizes
AR.
With fixed instruments and normal errors, the quadratic
forms in the numerator and denominator of (8) are indepen-
dent chi-squared random variables under the null hypothe-
sis, and AR0  has an exact FK T −K null distribution. Under
the more general conditions of weak-instrument asymptotics,
d
AR0 →5k2 /K under the null hypothesis, regardless of the
value of 2 /K. Thus the AR statistic provides a fully robust
test of the hypothesis  = 0 .
The AR statistic is profligate in its use of overidentifying
restrictions in the sense that the numerator projects y − Y0 on
Z rather than on a subspace of Z, leading to a loss of power
relative to the infeasible power envelope when  is overi-
dentified. Moreover, the AR statistic can reject either because
 = 0 or because the instrument orthogonality conditions
fail, so inference based on the AR statistic differs from infer-
ence based on conventional GMM test statistics, for which the
maintained hypothesis is that the instruments are valid. For
these reasons, other statistics have been proposed for testing
 = 0 with the aim of improving power relative to AR0 
when  is overidentified.
5.2.2. Kleibergen’s Statistic. Kleibergen (2001) pro-
posed the statistic

 2
 
which, following Moreira (2001), we have written in terms
of
 and . If K = 1, then K0  = AR0 . Kleibergen
showed that under either conventional or weak-instrument
asymptotics, K0  has a 512 null limiting distribution.
5.2.3. Moreira’s Statistic. Moreira (2002) proposed test-
ing  = 0 using the conditional likelihood ratio test statistic

1    
M0  =

− 
2
 
 
+ 

+  2 − 4 


   − 
 2  (10)
The (weak-instrument) asymptotic distribution of M0 
under the null, conditional on  = 7, is nonstandard and
depends on 0 and 7. Moreira (2002) suggested computing
the null distribution by Monte Carlo simulation.
5.3 Conservative Tests
Staiger and Stock (1997) suggested testing  = 0 using a
Bonferroni test. Wang and Zivot (1998) and Zivot, Startz, and
Nelson (1998) proposed a modification of conventional GMM
statistics in which u2 is estimated under the null hypothesis.
524 Journal of Business & Economic Statistics, October 2002

Under weak-instrument asymptotics, these tests are conserva- (a)

µ 2/K = 1 ρ = 0.5
tive (i.e., their size is less than their significance level for some 1.0

values of the parameters). Numerical analysis suggests that AR K-test

Power envelope
these tests tend to have lower power than the Gaussian similar 0.8

tests. M-test

0.6

Power
5.4 Power Comparisons 0.4

The asymptotic power functions of the AR, Kleibergen, and

Moreira tests depend on 2 /K,  [the correlation between u
0.2

and v in (1) and (2)], and K, as well as on the true value

0.0
of . We consider two values of 2 /K 8 2 /K = 1, which -10.0 -7.5 -5.0 -2.5 0.0 2.5 5.0 7.5 10.0
corresponds to very weak instruments (nearly unidentified), β– β0

and 2 /K = 5, which corresponds to moderately weak instru- (b)

µ 2/K = 1 ρ = 0.99
ments. The two values of  considered correspond to moderate 1.0

endogeneity ( = 5) and very strong endogeneity ( = 99, as

used in Fig. 1). 0.8

Figure 2 presents weak-instrument asymptotic power func-

tions for K = 5 instruments, so the degree of overidentification 0.6

Power
is 4. The power depends on  − 0 but not on 0 , so Figure 2
applies to general 0 . The shaded region is the area between 0.4

Moreira’s (2001) infeasible asymptotic Gaussian power enve-

lope and the power function of the AR test; the challenge for 0.2

newly proposed fully robust tests is to have power functions as

close to the top of this region as possible. When 2 /K = 1 and
0.0
-10.0 -7.5 -5.0 -2.5 0.0 2.5 5.0 7.5 10.0
 = 5, all tests have poor power for all values of the parame- β– β0
ter space—a reassuring result given how weak the instruments (c) µ 2/K = 5 ρ = 0.5
are; moreover, all tests have power functions that are far from 1.0
the infeasible power envelope. Notably, the power functions
do not increase monotonically in
 − 0
. When 2 /K = 5, 0.8
the M test (but not the K test) approaches the infeasible enve-
lope for both values of . 0.6
Power

Figure 3 presents the corresponding power functions for

many instruments (K = 50). In all cases, the M test is within 0.4

or toward the top of the shaded region; this is mainly (but not
always) the case for the K test, which has a power function 0.2

that, oddly, descends substantially for   0 . As Figure 3

makes clear, when K is large, the AR test has relatively 0.0

low power (arising from its inefficient use of overidentifying

-10.0 -7.5 -5.0 -2.5 0.0 2.5 5.0 7.5 10.0

β– β0
restrictions), and substantial power improvements are possi-
ble, particularly by using the M test. (d) µ 2/K = 5 ρ = 0.99
1.0

5.5 Robust Confidence Sets 0.8

Due to the duality between hypothesis tests and confidence 0.6

sets, these tests can be used to construct fully robust confi-
Power

dence sets. For example, a fully robust 95% confidence set 0.4
can be constructed as the set of 0 for which the AR statistic,
AR0 , fails to reject at the 5% significance level. In gen- 0.2
eral, this approach requires evaluating the test statistic for all
points in the parameter space, although for some statistics the 0.0
confidence interval can be obtained by solving a polynomial -10.0 -7.5 -5.0 -2.5 0.0 2.5 5.0 7.5 10.0

equation.
When the instruments are weak, these sets can have infinite
volume. For example, because the AR statistic is a ratio of
quadratics, it can have a finite maximum, and when 2 = 0,
any point in the parameter space will be contained in the AR
confidence set with probability 95%. This does not imply that
β– β0
Figure 2. Weak-Instrument Asymptotic Power of Gaussian Similar
Tests for K = 5 Instruments. The upper boundary of the shaded area
is the Gaussian power envelope, the lower boundary is the power of
the AR test. The other two power functions are for Kleibergen’s and
Moreira’s tests.
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM 525

(a) µ 2/K = 1 ρ = 0.5 these methods waste information or are unnecessarily impre-
1.0
cise; rather, if instruments are weak, then there simply is lim-
Power envelope
ited information to use to make inferences about . This point
0.8
was made formally by Dufour (1997), who showed that under
weak-instrument asymptotics, a confidence set for  must have
0.6 infinite expected volume if it is to have nonzero coverage uni-
Power

formly in the parameter space, as long as 2 is fixed and finite.

0.4 This infinite expected volume condition is shared by confi-
K-test
dence sets constructed using any of the fully robust methods
0.2 AR of this section (see Zivot et al. 1998 for further discussion).
M-test

0.0
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
6. PARTIALLY ROBUST INFERENCE
β– β0 WITH WEAK INSTRUMENTS
(b) 2
µ /K = 1 ρ = 0.99 Although the fully robust tests discussed in the previous
1.0 section always control size, they can be difficult to compute.
Moreover, for n > 1, they do not readily provide point esti-
0.8 mates, and confidence intervals for individual elements of 
must be obtained by conservative projection methods. The
0.6 methods described in this section are relatively easy to com-
Power

pute, and inference proceeds using conventional normal fixed-

0.4 model asymptotic approximations. These methods are partially
robust to weak instruments in the sense that they are more
0.2
reliable than TSLS when instruments are weak.

0.0 6.1 k-Class Estimators

-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
β– β0 The k-class estimator of  is k ˆ = Y  I −
−1 
kMZ Y Y I − kMZ y . This class includes TSLS (for
(c) µ 2/K = 5 ρ = 0.5
1.0 which k = 1), LIML, and some alternatives that improve on
TSLS when instruments are weak.
0.8 6.1.1. Limited-Information Maximum Likelihood. LIML
is a k-class estimator where k = kLIML is the smallest root of
0.6
the determinantal equation
Y  Y − kY  MZ Y
= 0. Although the
mean of the LIML estimator does not exist because its distri-
Power

0.4
bution has fat tails, its median is typically much closer to 
than is the mean or median of TSLS. In the fixed-instrument,
normal-error model, the bias of TSLS increases with K, but
0.2
the bias of LIML does not (Rothenberg 1984). When the
instruments are fixed and the errors are symmetrically dis-
tributed, LIML is the best median-unbiased k-class estima-
0.0
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
β– β0 tor to second order (Rothenberg 1983). Moreover, unlike
TSLS, LIML is consistent under many-instrument asymptotics
(d) µ 2/K = 5 ρ = 0.99
1.0
(Bekker 1994).
6.1.2. Fuller-k Estimator. Fuller (1977) proposed an
0.8
alternative k-class estimator that sets k = kLIML − b/T − K,
where b is a positive constant. With fixed instruments and nor-
mal errors, the Fuller-k estimator with b = 1 is best unbiased
0.6
to second order (Rothenberg 1984). In Monte Carlo simula-
Power

tions, Hahn et al. (2001a) reported substantial reductions in

bias and mean squared error (MSE) using Fuller-k estimators,
0.4

relative to TSLS and LIML, when instruments are weak.

0.2
6.1.3. Bias-Adjusted Two-Stage Least Squares. Donald
and Newey (2001) considered a bias-adjusted TSLS estima-
0.0 tor (BTSLS), a k-class estimator with k = T /T − K + 2,
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
β– β0
modifying an estimator previously proposed by Nagar (1959).
Rothenberg (1984) showed that BTSLS is unbiased to second
Figure 3. Weak-Instrument Asymptotic Power of Gaussian Similar order in the fixed-instrument, normal-error model. Donald and
Tests for K = 50 Instruments. See the legend to Figure 2. Newey provided expressions for the second-order asymptotic
526 Journal of Business & Economic Statistics, October 2002

MSE of BTSLS, TSLS, and LIML as a function of the number (a)

of instruments K. In Monte Carlo simulations, these authors 18

found that selecting the number of instruments to minimize 16

the second-order MSE generally improves performance. Chao JIVE
and Swanson (2001) derived the bias and MSE of TSLS under 14

weak-instrument asymptotics, modified to allow the number 12

F -statistic
of instruments to increase with the sample size. They reported TSLS
improvements in Monte Carlo simulations by incorporating 10

bias adjustments. 8
6.1.4. Jackknife Instrumental Variables. Angrist, Imbens, BTSLS
and Krueger (1999) proposed the jackknife instrumental vari- 6

ables estimator (JIVE), ˆ JIVE = Y
 y, where the ith

 Y −1 Y
4
row of Y −i and 

is Zi  −i is the estimator of  computed Fuller-k
using all but the ith observation. They showed that JIVE and 2
3 5 7 9 11 13 15 17 19
TSLS are asymptotically equivalent under conventional fixed- K
model asymptotics. Calculations drawing on work of Chao and (b)
Swanson (2002) reveal that under weak-instrument asymp- 18

totics, JIVE is asymptotically equivalent to a k-class estima- 16

tor with k = 1 + K/T − K. Theoretical calculations (Chao TSLS
and Swanson 2002) and Monte Carlo simulations (Angrist, 14

Imbens, and Krueger 1999; Blomquist and Dahlberg 1999)

12
indicate that JIVE improves on TSLS when there are many F -statistic
instruments. 10

6.2 Comparisons
One way to assess how robust an estimator or test is to weak
instruments is to characterize the size of its weak instrument
region. When n = 1, this can be done by computing the critical
value of the first-stage F statistic testing (at the 5% level) the
null hypothesis that 2 /K is too small to ensure a desired
degree of reliability under weak-instrument asymptotics (i.e.,
the instruments are weak) against the alternative that it exceeds
the threshold value of 2 /K (i.e., the instruments are strong).
This is the approach taken in Table 1 for TSLS, and Figure 4
applies it to the other estimators discussed in this section. In
Figure 4(a), the weak-instrument set is defined to be the set
of 2 /K such that the relative bias of the estimator exceeds
10%; in Figure 4(b), the weak-instrument set is instead defined
so that a nominal 5% test of  = 0 , based on the relevant
t statistic, can have size exceeding 15%. In the context of
Figure 4, the smaller the critical values, the more robust the
procedure.
As Figure 4 shows, LIML, BTSLS, JIVE, and the Fuller-k
estimator (with b = 1) generally have smaller critical values
than TSLS. In this sense, these four estimators are more robust
to weak instruments than TSLS. In contrast to TSLS, these
critical values decrease as a function of K. For K ≥ 10, the
critical values of the first-stage F statistic fall to 5 or less,
well below those for TSLS. In this sense, these partially robust
methods evidently provide relatively reliable alternatives in
applications with weak instruments.
7. GENERALIZED METHOD OF MOMENTS
INFERENCE IN GENERAL NONLINEAR MODELS
It has been recognized for some time that the usual large-
sample normal approximations to GMM statistics in general
nonlinear models can provide poor approximations to exact
8
JIVE BTSLS
6
4
LIML
2
1 3 5 7 9 11 13 15 17 19
K
Figure 4. Critical Values for Weak-Instrument Tests Based on the
First-Stage F Statistic for the TSLS, LIML, BTSLS, JIVE, and Fuller-k
Estimators As a Function of the Number of Instruments (K). The critical
value is for a 5% test of the null hypothesis that the instruments are
weak, defined as (a) the weak-instrument asymptotic relative bias of the
estimator exceeds 10% and (b) the weak-instrument asymptotic size of
the 5% Wald test can exceed 15%.
sampling distributions in problems of applied interest. For
example, Hansen et al. (1996) examined GMM estimators
of various intertemporal asset pricing models using a Monte
Carlo design calibrated to match U.S. data. They found that
in many cases, inferences based on the usual normal approxi-
mations are misleading (see also Tauchen 1986; Kocherlakota
1990; Ferson and Foerester 1994; Smith 1999). As discussed
in Section 3.2, weak instruments are a plausible source of
these problems.
The methods of Sections 4–6 apply to the linear IV model
with homoscedastic, serially uncorrelated errors. This section
provides a nontechnical discussion of methods that apply
when the errors are heteroscedastic or serially correlated
and/or when the model is nonlinear, that is, extensions of the
linear methods for iid data to general GMM. We begin by
briefly discussing the problems posed by weak instruments in
nonlinear GMM and suggest that a better term in this context
is weak identification. We then briefly survey the quite incom-
plete literature on detection of weak identification and on pro-
cedures that are fully or partially robust to weak identification.
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM 527

7.1 Weak Identification in Nonlinear GMM (Hansen et al. 1996) in which the weight matrix is evaluated
at the same parameter value as the numerator:
In GMM, the n × 1 parameter vector : is identified by the



G conditional mean restrictions EhYt  :0 
Zt
= 0, where :0 1 T
1 T
is the true value of : and Zt is a K-vector of instruments; this CU
ST : = 
< : W : −1
< :  (11)
T t=1 t T t=1 t
in turn implies E<t :0 
= 0, where <t : = hYt  : ⊗ Zt .
If the instruments are relevant, then EhYt  : ⊗ Zt
= 0 for
where W  : = T −1
Tt=1 <t : − <:
<
¯ ¯ 
t : − <:
and
: = :0 , a necessary condition for : to be identified. In the ¯ −1

T
linear model, weak instruments arise when EhYt  : ⊗ Zt
is <: = T t=1 <t :. [If <t : is serially correlated, then
W : is replaced by an estimator of the spectral density of
nearly 0 for : = :0 ; that is, when Zt is nearly uncorrelated with
the model error term even at false values of :. More generally, <t : at frequency 0.]
in nonlinear GMM, if EhYt  : ⊗ Zt
is nearly 0 for : = :0 , Under the null hypothesis : = :0 , STCU :0  is asymptoti-
2
then : can be thought of as being weakly identified. cally 5GK distributed, whether identification is weak or strong
(Stock and Wright 2000). If the instruments are relevant, then
Because there is no exact sampling theory for GMM esti-
under the alternative that : = :0 , the “numerator moments” of
mators, formal treatments of weak identification in GMM rely
STCU :0  have nonzero expectation. A confidence set for : is
on asymptotics. One approach is to use stochastic expansions
computed by inverting the STCU : statistic numerically (see
in orders of T 1/2 ; however, as in the linear case, the result-
Stock and Wright 2000; Ma 2002 for examples).
ing approximations seem likely to be poor when identification
7.3.2. Kleibergen’s Statistic. Kleibergen (2002) pro-
is very weak. A second approach (Stock and Wright 2000)
posed testing the hypothesis : = :0 using a generaliza-
is to use an asymptotic nesting in which, loosely speaking,
tion of K0  and showed that the proposed statistic has a
the GMM version of the concentration parameter is fixed as
5n2 distribution under both conventional asymptotics and the
T → . This yields a stochastic process representation of the
limiting objective function, which in the linear case simplifies
to the weak-instrument asymptotics discussed in Section 2.2.
7.2 Detecting Weak Identification
An implication of weak identification is that GMM estima-
tors can exhibit a variety of pathologies. For example, two-
step GMM estimators and iterated GMM point estimators can
be quite different and can produce quite different confidence
sets. If identification is weak, then GMM estimates can be
sensitive to the addition of instruments or to changes in the
sample. If these features are present in an empirical applica-
tion, then they can be symptomatic of weak identification.
The only formal test for weak identification in nonlinear
GMM of which we are aware is that proposed by Wright
(2001). In the conventional asymptotic theory of GMM, the
identification condition requires the gradient of <t :0  to have
full column rank. Wright (2001) proposed a test of the hypoth-
esis of a complete failure of this rank condition. Thus Wright’s
test, like Cragg and Donald’s (1993) in the linear model, is
strictly a test for nonidentification or underidentification, not
for weak identification.
7.3 Procedures That Are Fully Robust
to Weak Identification
We are aware of only two fully robust methods for test-
ing : = :0 in nonlinear GMM: a nonlinear AR statistic and
Kleibergen’s statistic.
7.3.1. Nonlinear Anderson–Rubin Statistic. Because the
numerator and denominator of the AR statistic (8) are eval-
uated at the true parameter value, it has a weak-instrument
weak-identification asymptotics of Stock and Wright (2000).
Kleibergen found in Monte Carlo simulations that his pro-
posed statistic generally gives a more powerful test than
STCU :0 , consistent with the improvement of the K test over
the AR test reported in Section 5.4.
7.4 Procedures That Are Partially Robust
to Weak Identification
Because there are estimators that improve on TSLS when
instruments are weak in the linear case, it stands to reason that
there should be estimators that improve on two-step GMM
in the nonlinear case. The limited work in this area to date
has yielded some promising results. Two GMM estimators
that appear to be partially robust to weak instruments are
the continuous-updating estimator (CUE) (Hansen et al. 1996)
and generalized empirical likelihood (GEL) estimator (Smith
1997). The CUE minimizes STCU : in (11). In the linear
model, the CUE is asymptotically equivalent to LIML under
weak-instrument and conventional asymptotics if the errors are
homoscedastic. GEL estimators represent a family of estima-
tors that contain empirical likelihood (Owen 1988; DiCiccio,
Hall, and Romano 1991), the CUE, and other estimators. The
GEL estimators have good properties in stochastic expansions
(Rothenberg 1999; Newey and Smith 2001). For example, all
GEL estimators are like LIML, BTSLS, JIVE, and the Fuller-
k estimator in the linear model, in the sense that their second-
order bias is less than that of the two-step GMM estimator.
Work on GEL estimators in the context of weak instruments is
promising but young; the reader is referred to Imbens (2002)
for further discussion.
8. CONCLUSIONS
asymptotic FK  distribution even if the unknown parame- Many of the extensions of GMM since Hansen’s (1982)
ters are poorly identified. This observation suggests tests of and Hansen and Singleton’s (1982) seminal work can be seen
: = :0 based on the nonlinear analog of the AR statistic, which as attempts to improve the performance of GMM in circum-
is the so-called continuous-updating GMM objective function stances of practical interest to empirical economists. One such
528
circumstance is the presence of weak instruments or weak
identification.
Despite the evolving nature of the literature, this survey sug-
gests that there are some useful methods that practitioners can
adopt to address concerns about weak instruments. In the lin-
ear IV model with homoscedastic errors and one endogenous
regressor, applied researchers should at least use the tools of
Section 4 to assess whether weak instruments potentially are
a problem in a given application, for example, by checking
the first-stage F statistic. If the first-stage F statistic is small,
say <10, and if the errors appear to be homoscedastic and
serially uncorrelated, then either a fully robust method (our
preference) from Section 5 or a partially robust method from
Section 6 can be used. Even if F > 10, it is prudent to check
the results using LIML, BTSLS, JIVE, or the Fuller-k estima-
tor, especially when the number of instruments is large. In the
GMM case (i.e., the moments are nonlinear in the parameters
and/or the errors are heteroscedastic or serially correlated),
then one or more of the methods of Sections 7.3 and 7.4 can
be used.
There are a number of related topics that, because of space
limitations, have not been discussed in this survey. Because
we have focused on weak instruments, we did not discuss the
problem of estimation when some instruments are strong and
others are weak. In that circumstance, one way to proceed is
to try to cull the weak instruments from the strong and to
use only the strong (see Hall and Inoue 2001; Hall and Peixe
2001; Donald and Newey 2001). A second omitted topic is
estimation of linear panel data models with a lagged depen-
dent variable, in which instruments (lags) are weak if the lag
coefficient is almost 1; recent work in this area includes that of
Kiviet (1995), Alonso-Borrego and Arellano (1996), and Hahn
et al. (2001b). A third omitted issue is combining weak instru-
ments with a failure of exogeneity restrictions (as emphasized
by Bound et al. 1995). On these and related topics, much work
remains.
ACKNOWLEDGMENTS
The authors thank Joshua Angrist, Whitney Newey, Adrian
Pagan, Marcelo Moreira, and Eric Swanson for comments
on an earlier draft. This research was supported in part by
National Science Foundation grant SBR-9730489.
[Received June 2002. Revised June 2002.]
REFERENCES
Alonso-Borrego, C., and Arellano, M. (1996), “Symmetrically Normalized
Instrumental-Variable Estimation Using Panel Data,” Journal of Business
& Economic Statistics, 17, 36–49.
Anderson, T. W. (1976), “Estimation of Linear Functional Relationships:
Approximate Distribution and Connections With Simultaneous Equations
in Econometrics,” Journal of the Royal Statistical Society, Ser. B, 38, 1–36.
Anderson, T. W., and Rubin, H. (1949), “Estimation of the Parameters of a
Single Equation in a Complete System of Stochastic Equations,” Annals of
Mathematical Statistics, 20, 46–63.
Angrist, J. D., Imbens, G. W., and Krueger, A. B. (1999), “Jackknife
Instrumental Variables Estimation,” Journal of Applied Econometrics, 14,
57–67.
Angrist, J. D., and Krueger, A. B. (1991), “Does Compulsory School Atten-
dance Affect Schooling and Earnings,” Quarterly Journal of Economics,
106, 979–1014.
Journal of Business & Economic Statistics, October 2002
Bekker, P. A. (1994), “Alternative Approximations to the Distribution of
Instrumental Variables Estimators,” Econometrica, 62, 657–681.
Bekker, P. A., and van der Ploeg, J. (1999), “Instrumental Variable Estimation
Based on Grouped Data,” manuscript, University of Groningen, Dept. of
Economics.
Blomquist, S., and Dahlberg, M. (1999), “Small Sample Properties of LIML
and Jackknife IV Estimators: Experiments With Weak Instruments,” Jour-
nal of Applied Econometrics, 14, 69–88.
Bound, J., Jaeger, D. A., and Baker, R. (1995), “Problems With Instrumental
Variables Estimation When the Correlation Between the Instruments and
the Endogenous Explanatory Variables Is Weak,” Journal of the American
Statistical Association, 90, 443–450.
Buse, A. (1992), “The Bias of Instrumental Variable Estimators,” Economet-
rica, 60, 173–180.
Campbell, J. Y. (2001), “Consumption-Based Asset Pricing,” in Handbook of
the Economics of Finance, forthcoming.
Campbell, J. Y., and Mankiw, N. G. (1989), “Consumption, Income and
Interest Rates: Reinterpreting the Time Series Evidence,” NBER Macroe-
conomics Annual, 4, 185–216.
Chao, J., and Swanson, N. R. (2001), “Bias and MSE Analysis of the IV
Estimator Under Weak Identification With Application to Bias Correction,”
unpublished manuscript, Purdue University.
(2002), “Consistent Estimation With a Large Number of Weak Instru-
ments,” unpublished manuscript, Purdue University.
Choi, I., and Phillips, P. C. B. (1992), “Asymptotic and Finite Sample Distri-
bution Theory for IV Estimators and Tests in Partially Identified Structural
Equations,” Journal of Econometrics, 51, 113–150.
Cragg, J. G., and Donald, S. G. (1993), “Testing Identifiability and Specifica-
tion in Instrumental Variable Models,” Econometric Theory, 9, 222–240.
DiCiccio, T., Hall, P., and Romano, J. (1991), “Empirical Likelihood is
Bartlett-Correctable,” The Annals of Statistics, 19, 1053–1061.
Donald, S. G., and Newey, W. K. (2001), “Choosing the Number of Instru-
ments,” Econometrica, 69, 1161–1191.
Dufour, J. M. (1997), “Some Impossibility Theorems in Econometrics With
Applications to Structural and Dynamic Models,” Econometrica, 65, 1365–
1387.
Ferson, W. E., and Foerester, S. R. (1994), “Finite Sample Properties of the
Generalized Method of Moments in Tests of Conditional Asset Pricing
Models,” Journal of Financial Economics, 36, 29–55.
Fuhrer, J. C., and Moore, G. R. (1995), “Inflation Persistence,” Quarterly
Journal of Economics, 110, 127–159.
Fuhrer, J. C., and Rudebusch, G. D. (2002), “Estimating the Euler Equation
for Output,” unpublished manuscript, Federal Reserve Bank of Boston.
Fuller, W. (1977), “Some Properties of a Modification of the Limited Infor-
mation Estimator,” Econometrica, 45, 939–953.
Gali, J., and Gertler, M. (1999), “Inflation Dynamics: A Structural Economet-
ric Analysis,” Journal of Monetary Economics, 44, 195–222.
Hahn, J., and Hausman, J. (2002), “A New Specification Test for the Validity
of Instrumental Variables,” Econometrica, 70, 163–189.
Hahn, J., Hausman, J., and Kuersteiner, G. (2001a), “Higher Order MSE of
Jackknife 2SLS,” unpublished manuscript, Massachusetts Institute of Tech-
nology, Dept. of Economics.
(2001b), “Bias Corrected Instrumental Variables Estimation for
Dynamic Panel Models With Fixed Effects,” unpublished manuscript,
Massachusetts Institute of Technology, Dept. of Economics.
Hall, A. R., and Inoue, A. (2001), “A Canonical Correlations Interpretation of
Generalized Method of Moments Estimation With Applications to Moment
Selection,” unpublished manuscript, North Carolina State University.
Hall, A. R., and Peixe, F. P. M. (2001), “A Consistent Method for the Selection
of Relevant Instruments,” unpublished manuscript, North Carolina State
University.
Hall, A. R., Rudebusch, G. D., and Wilcox, D. W. (1996), “Judging Instrument
Relevance in Instrumental Variables Estimation,” International Economic
Review, 37, 283–289.
Hall, R. E. (1988), “Intertemporal Substitution in Consumption,” Journal of
Political Economy, 96, 339–357.
Hansen, L. P. (1982), “Large Sample Properties of Generalized Method of
Moments Estimators,” Econometrica, 50, 1029–1054.
Hansen, L. P., Heaton, J., and Yaron, A. (1996), “Finite Sample Properties
of Some Alternative GMM Estimators,” Journal of Business & Economic
Statistics, 14, 262–280.
Hansen, L. P., and Singleton, K. J. (1982), “Generalized Instrumental Vari-
ables Estimation of Nonlinear Rational Expectations Models,” Economet-
rica, 50, 1269–1286.
(1983), “Stochastic Consumption, Risk Aversion and the Tem-
poral Behavior of Asset Returns,” Journal of Political Economy, 91,
249–265.
Imbens, G. W. (2002), “Generalized Method of Moments and Empirical Like-
lihood,” Journal of Business & Economic Statistics, forthcoming.
Stock, Wright, and Yogo: Weak Instruments and Identification in GMM
Kiviet, J. F. (1995), “On Bias, Inconsistency, and Efficiency of Various Esti-
mators in Dynamic Panel Data Models,” Journal of Econometrics, 68,
1–268.
Kleibergen, F. (2001), “Pivotal Statistics for Testing Structural Parameters in
Instrumental Variables Regression,” Econometrica, forthcoming.
(2002), “Testing Parameters in GMM Without Assuming That They
Are Identified,” unpublished manuscript, University of Amsterdam, Dept.
of Economics.
Kocherlakota, N. (1990), “On Tests of Representative Consumer Asset Pricing
Models,” Journal of Monetary Economics, 26, 285–304.
Kunitomo, N. (1980), “Asymptotic Expansions of the Distributions of Esti-
mators in a Linear Functional Relationship and Simultaneous Equations,”
Journal of the American Statistical Association, 75, 693–700.
Ma, A. (2002), “GMM Estimation of the New Phillips Curve,” Economics
Letters, 76, 411–417.
Mavroeidis, S. (2001), “Identification and Misspecification Issues in Forward
Looking Monetary Models,” unpublished manuscript, Oxford University,
Dept. of Economics.
Moreira, M. J. (2001), “Tests with Correct Size When Instruments Can
Be Arbitrarily Weak,” unpublished manuscript, University of California
Berkeley, Dept. of Economics.
(2002), “A Conditional Likelihood Ratio Test for Structural Models,”
unpublished manuscript, University of California Berkeley, Dept. of Eco-
nomics.
Morimune, K. (1983), “Approximate Distributions of k-class Estimators When
the Degree of Overidentifiability Is Large Compared With the Sample
Size,” Econometrica, 51, 821–841.
Nagar, A. L. (1959), “The Bias and Moment Matrix of the General k-Class
Estimators of the Parameters in Simultaneous Equations,” Econometrica,
27, 575–595.
Neely, C. J., Roy, A., and Whiteman, C. H. (2001), “Risk Aversion Versus
Intertemporal Substitution: A Case Study of Identification Failure in the
Intertemporal Consumption Capital Asset Pricing Model,” Journal of Busi-
ness & Economic Statistics, 19, 395–403.
Nelson, C. R., and Startz, R. (1990a), “Some Further Results on the Exact
Small Sample Properties of the Instrumental Variables Estimator,” Econo-
metrica, 58, 967–976.
(1990b), “The Distribution of the Instrumental Variable Estimator and
Its t Ratio When the Instrument Is a Poor One,” Journal of Business, 63,
S125–S140.
Newey, W. K., and Smith, R. J. (2001), “Higher Order Properties of GMM and
Generalized Empirical Likelihood Estimators,” unpublished manuscript,
Massachusetts Institute of Technology, Dept. of Economics.
Owen, A. (1988), “Empirical Likelihood Ratios Confidence Intervals for a
Single Functional,” Biometrika, 75, 237–249.
Pagan, A. R., and Robertson, J. C. (1998), “Structural Models of the Liquidity
Effect,” The Review of Economics and Statistics, 80, 202–217.
529
Phillips, P. C. B. (1984), “Exact Small Sample Theory in the Simultaneous
Equations Model,” in Handbook of Econometrics, Vol. 1, eds. Z. Griliches
and M. D. Intriligator, Amsterdam: North-Holland.
Richardson, D. H. (1968), “The Exact Distribution of a Structural Coef-
ficient Estimator,” Journal of the American Statistical Association, 63,
1214–1226.
Rothenberg, T. J. (1983), “Asymptotic Properties of Some Estimators in Struc-
tural Models,” in Studies in Econometrics, Time Series, and Multivariate
Statistics, eds. S. Karlin, T. Amemiya, and L. A. Goodman, New York:
Academic Press.
(1984), “Approximating the Distribution of Econometric Estimators
and Test Statistics,” in Handbook of Econometrics, Vol. 2, eds. Z. Griliches
and M. D. Intriligator, Amsterdam: North-Holland.
(1999), “Higher Order Properties of Empirical Likelihood for Simulta-
neous Equations,” unpublished manuscript, University of California Berke-
ley, Dept. of Economics.
Sawa, T. (1969), “The Exact Sampling Distribution of Ordinary Least Squares
and Two-Stage Least Squares Estimators,” Journal of the American Statis-
tical Association, 64, 923–936.
Shea, J. (1997), “Instrument Relevance in Multivariate Linear Models: A Sim-
ple Measure,” The Review of Economics and Statistics, 79, 348–352.
Sims, C. A. (1980), “Macroeconomics and Reality,” Econometrica, 48, 1–48.
Smith, D. C. (1999), “Finite Sample Properties of Tests of the Epstein-Zin
Asset Pricing Model,” Journal of Econometrics, 93, 113–148.
Smith, R. (1997), “Alternative Semiparametric Likelihood Approaches to
Generalized Method of Moments Estimation,” Economic Journal, 107,
503–519.
Staiger, D., and Stock, J. H. (1997), “Instrumental Variables Regression With
Weak Instruments,” Econometrica, 65, 557–586.
Stock, J. H., and Wright, J. H. (2000), “GMM With Weak Identification,”
Econometrica, 68, 1055–1096.
Stock, J. H., and Yogo, M. (2001), “Testing for Weak Instruments in Linear
IV Regression,” unpublished manuscript, Harvard University.
Tauchen, G. (1986), “Statistical Properties of Generalized Method of
Moments Estimators of Structural Parameters Obtained From Financial
Market Data,” Journal of Business & Economic Statistics, 4, 397–425.
Wang, J., and Zivot, E. (1998), “Inference on Structural Parameters in Instru-
mental Variables Regression With Weak Instruments,” Econometrica, 66,
1389–1404.
Wright, J. H. (2001), “Detecting Lack of Identification in GMM,” Economet-
ric Theory, forthcoming.
Yogo, M. (2002), “Estimating the Elasticity of Intertemporal Substitution
When Instruments Are Weak,” unpublished manuscript, Harvard Univer-
sity, Dept. of Economics.
Zivot, E., Startz, R., and Nelson, C. R. (1998), “Valid Confidence Intervals and
Inference in the Presence of Weak Instruments,” International Economic
Review, 39, 1119–1246.